Joint Efficacy and Safety Modeling for a Dose-Ranging Trial

Why model efficacy and safety together?

A dose-ranging trial can show that increasing dose improves efficacy while also increasing adverse effects. Considering each endpoint in isolation may therefore give an incomplete basis for comparing regimens. A joint analysis represents both trajectories and their dependence, so uncertainty in a combined decision quantity reflects how efficacy and safety vary together. Joint modeling does not, by itself, define what balance is acceptable; that requires a clinical decision rule specified in context.

Consider a hypothetical randomized trial with placebo and doses of 5, 10, and 20 mg. Each participant has an efficacy score and a safety-burden score measured at weeks 2, 4, and 8. Efficacy is expressed as improvement from baseline, so larger positive values are better. Safety burden is expressed so larger positive values indicate more adverse effects. These two scales have different units and meanings; any combined score must state how it weights them.

Model: Let \(d_i\) be participant \(i\)'s assigned dose in mg, and let \(t_j\) be a measurement week. One illustrative bivariate longitudinal model is

\begin{aligned} Y_{E,ij} &= \alpha_{E,j} + E_{\max}\frac{d_i}{ED_{50}+d_i}\frac{t_j}{8} + b_{E,i}+\varepsilon_{E,ij},\\ Y_{S,ij} &= \alpha_{S,j} + \beta d_i\frac{t_j}{8} + b_{S,i}+\varepsilon_{S,ij}. \end{aligned}

Here \(Y_{E,ij}\) and \(Y_{S,ij}\) are efficacy improvement and safety burden, respectively; \(\alpha_{E,j}\) and \(\alpha_{S,j}\) are time-specific mean changes under placebo; \(E_{\max}\) and \(ED_{50}\) describe the saturating efficacy dose-response; and \(\beta\) is the dose-related safety slope. The factors \(t_j/8\) specify a linear increase in each drug effect through week 8. The paired random intercepts \((b_{E,i},b_{S,i})\) have a bivariate normal distribution with mean zero and covariance matrix \(G\). At each visit, the paired residuals \((\varepsilon_{E,ij},\varepsilon_{S,ij})\) have mean zero and covariance matrix \(R\), and residual pairs are independent across visits and participants. Random effects and residuals are mutually independent.

This model is a useful starting point, not a universal form. It assumes the chosen mean curves are adequate, the random effects and residuals have the stated distributions, and the randomization and follow-up support comparisons across doses. The random intercepts capture persistent participant-level departures from the mean, while the off-diagonal elements of \(G\) and \(R\) allow efficacy and safety to co-vary. The model can be fitted by maximum likelihood or restricted maximum likelihood, with nonlinear optimization for the \(E_{\max}\) curve. If observations are missing, likelihood-based inference also relies on a missing-at-random assumption conditional on variables in the model; informative dropout may require additional modeling.

Fitting and interpreting the longitudinal dose response

For illustration, suppose the week-8 observed group means are as follows. These are hypothetical summaries; a model fit would ordinarily use participant-level repeated measurements, not just these means.

Assigned dose (mg)	Week-8 efficacy improvement (points)	Week-8 safety burden (points)
0	0.1	0.0
5	4.7	0.7
10	6.5	1.1
20	8.8	2.5

Suppose the fitted efficacy parameters are \(E_{\max}=12.0\) points and \(ED_{50}=8.0\) mg, and the fitted safety slope is \(\beta=0.120\) safety-burden points per mg. The estimates summarize the overall repeated data, including weeks 2 and 4, rather than being obtained by forcing the model through the four displayed week-8 means. The estimated placebo time effects account for change that is not attributed to dose.

Worked Example: Predicting both outcomes over time

Calculate the model-predicted drug effects at 10 mg at weeks 4 and 8. For simplicity, compare with the placebo trajectory at the same visit, so the \(\alpha\) terms cancel.

At week 8, the efficacy effect is

12.0\frac{10}{8.0+10}\frac{8}{8} =12.0\left(\frac{10}{18}\right) =6.667\text{ efficacy points}.

At week 4, the time factor is \(4/8=0.5\), so the efficacy effect is \(6.667(0.5)=3.333\) points. The safety effects are \(0.120(10)(4/8)=0.600\) points at week 4 and \(0.120(10)(8/8)=1.200\) points at week 8. Thus the fitted mean predicts increasing benefit and increasing safety burden over this interval. These are population mean contrasts with placebo, not predictions for every participant.

A linear time factor is a strong assumption: it rules out an early plateau or delayed onset in the drug effect. Plotting observed group trajectories and model-predicted trajectories by dose is therefore essential. Systematic curvature, dose-specific departures, or residual trends over time suggest that the mean structure needs revision. The assumed random-intercept structure should also be checked: if residual correlation across visits remains after accounting for random intercepts, a richer within-participant covariance structure may be needed.

What the joint covariance adds

A separate efficacy model and safety model can estimate each endpoint's mean curve. But two separate fits do not, on their own, estimate the sampling covariance between the two predictions. That covariance matters whenever a decision uses both outcomes. In a joint fit, the estimated dependence can arise from shared participant-level random effects and from paired residuals. The covariance is not automatically positive: its sign and magnitude should be estimated and checked rather than assumed.

Key result: If \(\widehat E\) and \(\widehat S\) are estimated mean effects with variances \(V_E\) and \(V_S\), and covariance \(C_{ES}\), then for a weighted net-benefit measure \(U=\widehat E-\lambda\widehat S\),

\operatorname{Var}(U)=V_E+\lambda^2V_S-2\lambda C_{ES}.

The weight \(\lambda\) expresses the decision maker's trade-off in the endpoint units; it is not a parameter supplied by the statistical model.

The formula follows by expanding the variance of a difference. For any random quantities \(A\) and \(B\), \(\operatorname{Var}(A-B)=\operatorname{Var}(A)+\operatorname{Var}(B)-2\operatorname{Cov}(A,B)\). Set \(A=\widehat E\) and \(B=\lambda\widehat S\). Since \(\operatorname{Var}(\lambda\widehat S)=\lambda^2V_S\) and \(\operatorname{Cov}(\widehat E,\lambda\widehat S)=\lambda C_{ES}\), the displayed expression follows. Ignoring covariance is equivalent to setting \(C_{ES}=0\), an assumption that may be inconsistent with the fitted data.

Worked Example: Quantifying uncertainty in net benefit

Suppose the joint fit at 10 mg and week 8 estimates \(\widehat E=6.667\) points and \(\widehat S=1.200\) points. For a prespecified illustrative trade-off of \(\lambda=2.00\) efficacy points per safety-burden point, the estimated net benefit is

\widehat U=6.667-2.00(1.200)=4.267\text{ efficacy-equivalent points}.

Assume the fitted sampling variances are \(V_E=0.640\) points squared and \(V_S=0.0400\) safety-burden points squared, with covariance \(C_{ES}=0.0800\) efficacy-point times safety-burden-point. Then

\begin{aligned} \operatorname{Var}(\widehat U) &=0.640+(2.00)^2(0.0400)-2(2.00)(0.0800)\\ &=0.640+0.160-0.320=0.480,\\ SE(\widehat U)&=\sqrt{0.480}=0.6928\text{ efficacy-equivalent points}. \end{aligned}

Under an approximately normal sampling distribution, an illustrative 95% Wald interval for the mean net benefit is \(4.267\pm1.96(0.6928)\), or \(2.909\) to \(5.625\) efficacy-equivalent points. The interval describes uncertainty in the estimated mean net benefit under the model, not the range of outcomes expected for individual participants. If the two separate fits were combined while incorrectly setting covariance to zero, the variance would be \(0.640+0.160=0.800\), giving \(SE=0.8944\). In this example, positive covariance reduces uncertainty in the efficacy-minus-weighted-safety contrast; with a different sign or trade-off, covariance could instead increase it.

Using predictions to compare candidate regimens

A net-benefit score is one way to compare regimens, but it should not replace a clinically meaningful safety limit. A decision rule can, for example, exclude doses whose upper uncertainty bound for mean safety burden exceeds a prespecified threshold, then compare net benefit among the remaining doses. The threshold and the weight must be justified clinically before interpreting the ranking. Changing either can change the preferred regimen.

Worked Example: Selecting among doses with a safety constraint

Use the fitted week-8 mean effects from the model at doses 5, 10, and 20 mg. The efficacy predictions are \(12d/(8+d)\), and the safety predictions are \(0.120d\). Take \(\lambda=2.00\), and suppose the joint fit gives safety-prediction standard errors of 0.180, 0.200, and 0.250 points at 5, 10, and 20 mg, respectively. As an illustrative rule, retain doses only if the upper 95% normal bound for mean safety burden is below 2.00 points.

At 5 mg, efficacy is \(12(5)/(8+5)=4.615\) points and safety burden is \(0.120(5)=0.600\) points. Net benefit is \(4.615-2(0.600)=3.415\) points, and the upper safety bound is \(0.600+1.96(0.180)=0.9528\) points.

At 10 mg, efficacy is 6.667 points and safety burden is 1.200 points. Net benefit is \(6.667-2(1.200)=4.267\) points, and the upper safety bound is \(1.200+1.96(0.200)=1.592\) points.

At 20 mg, efficacy is \(12(20)/(8+20)=8.571\) points and safety burden is \(0.120(20)=2.400\) points. Net benefit is \(8.571-2(2.400)=3.771\) points, but its upper safety bound is \(2.400+1.96(0.250)=2.890\) points.

The 20 mg regimen fails the stated safety constraint. Both 5 mg and 10 mg satisfy it, and 10 mg has the larger estimated net benefit. Under this hypothetical decision rule, 10 mg is the preferred candidate. This is a model-based comparison, not proof that the dose is optimal in a future population. The estimated uncertainty and the prespecified trade-off should be reported alongside the choice.

Diagnostics, alternatives, and limitations

Joint modeling is most useful when the scientific question involves both endpoints or their dependence. For endpoint-specific description, two separate longitudinal models may be adequate and simpler. A joint model is preferable when estimating covariance-aware contrasts, combined scores, or uncertainty for a decision using both outcomes. In a balanced setting, endpoint-specific mean estimates from separate models may resemble those from a joint model; the key distinction is that separate fits do not supply the cross-endpoint covariance needed for valid joint uncertainty calculations.

Diagnostics should examine each endpoint and their pairing. Plot residuals against fitted values, dose, and time; inspect observed and fitted trajectories; and assess whether residual variability changes across doses or visits. Compare observed within-participant correlations with those implied by the fitted covariance structure. Boundary estimates for random-effect variances, unstable correlations, convergence warnings, or markedly different estimates across plausible starting values indicate that the covariance model may be too complex for the data. A poorly fitting mean curve can also make covariance estimates misleading.

The model's limitations are consequential. The chosen \(E_{\max}\) and linear safety forms may not represent the true dose-response, only four assigned dose levels provide limited information about curve shape, and a normal random-effects model can be sensitive to outliers. The illustrative time scaling imposes a particular trajectory. Furthermore, a net-benefit weight compresses two outcomes into one number and can obscure important endpoint-specific concerns. Report efficacy and safety predictions separately, the joint uncertainty used for the decision, and sensitivity to alternative reasonable weights or safety thresholds.

Common mistake: treating two marginal confidence intervals as a joint analysis. Separate intervals do not encode how the estimated outcomes co-vary. Estimate or otherwise justify their covariance before calculating uncertainty for a combined contrast.
Common mistake: interpreting correlation as a causal mechanism. A fitted association between efficacy and safety does not establish that improvement causes adverse effects, or vice versa.
Common mistake: choosing the dose with the largest efficacy estimate alone. A higher response can coincide with unacceptable safety burden. Apply the stated benefit-risk rule and present endpoint-specific results.
Common mistake: confusing mean uncertainty with individual variability. A confidence interval for a mean prediction is not a prediction interval for a participant's outcome; the latter also includes between-person and residual variation.

Check Your Understanding

Use the stated hypothetical model and decision-rule ideas. Solutions follow each question.

At 5 mg and week 8, calculate the predicted efficacy and safety effects. Solution: Efficacy is \(12(5)/(8+5)=4.615\) points; safety burden is \(0.120(5)=0.600\) points.
If the estimated efficacy-safety covariance is zero, what is the net-benefit variance at \(\lambda=2\) using \(V_E=0.640\) and \(V_S=0.0400\)? Solution: \(0.640+2^2(0.0400)-2(2)(0)=0.800\) efficacy-equivalent points squared.
Why can a joint model be necessary even when two separate models provide reasonable mean curves? Solution: The separate fits do not by themselves estimate cross-endpoint covariance, which is needed to calculate uncertainty for a quantity combining efficacy and safety.
What does a positive covariance do to the variance of \(\widehat E-2\widehat S\)? Solution: It reduces the variance through the term \(-2(2)C_{ES}\), provided the covariance is positive; the effect depends on the contrast's direction and weight.
Does a 95% interval for mean net benefit describe the likely range of individual participants' net outcomes? Solution: No. It quantifies uncertainty in the estimated mean under the model; individual prediction requires adding between-participant and residual variability.